autoscaler watches over Kue queue and scales k8s deployments off of pending jobs.
See charts/autoscaler.
Create the CRD and use the example AutoScalerWatchers:
$ kubectl create -f ./deploy
customresourcedefinition.apiextensions.k8s.io/autoscalerwatchers.tritonjs.com created
autoscalerwatcher.tritonjs.com/triton-downloader created
autoscalerwatcher.tritonjs.com/triton-converter createdIf you're running the rest of the triton platform, it will now automatically scale converter and downloader instances when jobs are available in the queue!
The autoscaler reads configuration from Kubernetes via the form of a CRD (Custom Resource Definition). This is used to configure the autoscaler on runtime, and dynamically configure watchers.
A watcher is a configuration of a queue that the autoscaler should listen on, currently only Kue is supported, but RabbitMQ and/or other queue technologies may be added in the future in the form of a adapters. These watchers create and a execute code based on Jobs.
A Job is used to tell the autoscaler what to do. These jobs are different from normal jobs because they, by default, are not active until a certain amount of time has passed and conditions are true for that entire time. (promote_after_minutes, and timeout_seconds). After promote_after_minutes passes, and timeout_seconds was never violated on updated_at, the job is automatically promoted to ready and the autoscaler then executes the operation.
Operations are what the autoscaler does when a Job is ready. Currently only two operations are supported: scaleUp and scaleDown. These modify kubernetes resources accordingly.
MIT